Machine Learning-based Energy Consumption models for Battery Electric Trucks

Abstract

Efforts to transition the heavy-duty vehicle sector from conventional diesel trucks to battery electric trucks (BETs) have generated vast interest in understanding the energy demand of heavy-duty BETs. Microscopic models are commonly used for modeling vehicle energy consumption during powertrain design and evaluation as they are very accurate. However, these models involve a number of input parameters and require high-resolution data for many of those inputs (e.g., second-by-second speed, acceleration, and road grade profiles). This level of input data may not be readily available, for example, when modeling energy consumption of multiple vehicles or for a large number of drive cycles. Mesoscopic models offer a practical alternative as the inputs needed, such as average traffic speed and road grade on a link-by-link basis, can be obtained more easily. This paper presents the development of mesoscopic energy consumption models for BETs using both real-world and microscopically simulated BET energy consumption datasets. A machine learning technique called random forest (RF) regressor was applied to the datasets to fit models. The results show that the RF regressor outperforms the classical linear regressor as evidenced by the resulting models having higher R2 values. When applied to the simulated dataset, the RF regressor can capture the behaviors of BET energy consumption well, where the R2 values of the resulting models are 0.86-0.89. When applied to the real-world dataset, the R2 values of the resulting models are only 0.50-0.52 as a large portion of the variance in the real-world dataset (e.g., cargo weight) is not captured.